Contact-Rich Robot Manipulation: How to Audit Force, Precision, Adaptation, and Recovery

The short answer

A contact-rich robot result should report what happened at the interface, not only whether the final object looked installed. The key evidence is a connected record of geometry, force, timing, adaptation, intervention, recovery, and repeat trials.

Vision can guide a gripper toward a slot. The difficult part begins when a small position or angle error turns into lateral force, friction, jamming, bent pins, a false seat, or a dropped part. Two actions can appear nearly identical in camera images while producing opposite physical outcomes.

Two September 2026 preprints make complementary parts of this evaluation problem visible:

• Facet-0 predicts the wrist force and torque expected after a proposed action, then uses deployment rollouts to value contact outcomes and adapt a bounded local action policy. Its authors report an 82% mean task success rate across five electronics assembly tasks, compared with 15% for the strongest matched baseline.
• Peg-in-Bench proposes modular, 3D-printable insertion parts and a seeded scenario generator. It varies geometry, clearance, position, orientation, and assembly structure so that a policy cannot be judged on one familiar peg and hole.

The first is a policy, dataset, and author-run physical evaluation. The second is a physical benchmark design. Neither is a production deployment, an independent replication, or a safety certification.

A useful contact-rich scorecard has six panels:

Panel	Minimum evidence
Task	part, geometry, sequence, initial-state distribution, success condition
Contact	clearance, force or torque sensing, peak off-axis force, configured limits, violations
Control	action space, compliant-control layer, observation-to-command latency, control rates
Learning	demonstrations, deployment episodes, interventions, resets, updated parameters, wall-clock cost
Recovery	detection rule, autonomous retry, human takeover, reset, recovered failures, damaged parts
Reproducibility	trial-level outcomes, physical artifacts, scenario seeds, code, checkpoints, versions, outside reproduction

This extends our diagnostic robot benchmark guide. That guide asks whether a benchmark required its intended observations and which constituent skill failed. This guide asks what must be measured when contact itself determines the outcome.

Why the final millimeter is a different control problem

Free-space motion and contact motion expose different failure modes.

During free-space approach, a policy can often correct a visible pose error without touching the environment. At contact, the robot and object become one coupled mechanical system. Motion depends on:

• position and orientation error;
• hole clearance and part geometry;
• friction and surface finish;
• material stiffness and compliance;
• end-effector and fixture rigidity;
• control frequency and delay;
• force or tactile sensing;
• commanded velocity and acceleration;
• controller gains and safety limits;
• whether the robot can detect and reverse a jam.

A camera may lose the relevant interface after the gripper occludes it. Wrist force and torque can reveal that displacement stopped while lateral force rose, but the signal arrives after contact begins. Tactile sensors may localize contact at the gripper, while a wrist sensor measures the combined wrench transmitted through the tool. Neither sensor automatically tells the policy which correction will work.

That creates four separate questions:

1. Observation: did the system detect the relevant contact state?
2. Prediction: did it anticipate the physical consequence of a proposed action?
3. Control: did the execution layer regulate motion and force appropriately?
4. Recovery: did it recognize failure and return to a valid state without hiding human help?

A paper can answer one question without answering the others.

What Facet-0 changes

Facet-0 begins with a PaliGemma vision-language backbone and a flow-matching action expert. The evaluated observation includes three synchronized RGB views, a task instruction, end-effector pose, gripper opening, and a ten-sample history of six-axis wrist force and torque.

The model jointly generates:

• a chunk of Cartesian pose and gripper commands;
• the future wrist-wrench profile expected after those commands.

The wrench output is predictive and non-commanded. Only the Cartesian action is sent toward execution. This distinction matters because “the model predicts force” and “the controller commands force” are different claims.

The paper describes a three-rate execution hierarchy:

Layer	Reported rate	Role
Coarse action expert	5 to 10 Hz	broad motion proposal at about 5 mm resolution
Contact refinement expert	20 Hz	action-wrench refinement at about 0.5 mm resolution
Compliant controller	200 Hz	closes the force loop on measured wrist force and torque

A task safety operator also clips targets to the workspace and limits per-step motion. The authors explicitly say this is a kinematic constraint, not a certified hard force bound or a passivity guarantee.

Why predict wrench rather than only read it?

Current wrench is a record of contact that earlier actions already produced. A behavior-cloning policy may receive that channel but learn to rely mostly on visual features. Facet-0 makes future wrench a supervised consequence of the proposed action. Its critic can then assign different values to two motions that make similar geometric progress but are expected to produce a clean insertion or a jam.

This is an architecture claim backed by the paper’s controlled recipe and ablations. It does not establish that future-wrench prediction is always necessary, that it replaces classical force control, or that the predicted wrench is calibrated outside the evaluated tasks.

The ManuFacet-1K training record

The authors describe ManuFacet-1K as approximately 1,000 hours of force-synchronized demonstrations and closed-loop rollouts. It covers CPU, RAM, disk, and GPU installation across three robot embodiments, two chassis families, multiple manufacturing cells, and a common sensing and teleoperation specification.

Each cell uses:

• a torque-controlled arm;
• a parallel gripper;
• a wrist six-axis force and torque sensor;
• three RGB cameras;
• joint proprioception;
• kinesthetic teleoperation with force feedback.

The training timeline is synchronized at 15 Hz, with higher-rate state and wrench streams retained for the 200 Hz controller. The curation process removes segment-level teleoperation faults, truncated chunks, sensor saturation, excessive calibration drift, and sustained occlusion. It retains corrected trajectories alongside anomalous trajectories so failure and recovery remain available for post-training.

The scale is important, but “1,000 hours” should not be read as 1,000 hours of autonomous successful assembly. The corpus combines demonstrations, policy rollouts, human interventions, free-space motion, contact phases, and recovery. Those categories have different evidentiary value.

The public project page, model page, dataset page, and code repository were reachable during our review. Availability of artifacts improves inspection. It does not by itself establish successful outside reproduction.

Reconstructing the 82% result

The controlled evaluation uses five electronics assembly tasks on a shared chassis fixture. The four insertion chains use clearances from 0.10 mm to 0.30 mm. A fifth task closes a CPU socket lever through two contact-critical presses.

Each method-task cell contains 20 trials, declared in advance with no reruns. Chassis pose is randomized per trial and an automated feeder supplies parts. A task succeeds only if the full sequence completes without a configured force violation.

Task	Tightest reported clearance	Facet-0 full system	Strongest baseline in that task
RAM installation	0.18 mm	95%	35%
CPU installation	0.10 mm	85%	20%
Disk installation	0.30 mm	95%	30%
GPU installation	0.18 mm	85%	10%
Lever closure	0.10 mm	50%	5%
Unweighted mean	not applicable	82%	15% for the strongest overall matched baseline

The baseline columns include π0.5, a version with force input, a RECAP-style reimplementation, GR00T N1.7, and a torque-aware VLA reimplementation. The paper states that baselines use the same dataset and matched budgets, but several are aligned reimplementations rather than untouched upstream releases. The comparison is therefore an author-run matched study, not a public cross-lab leaderboard.

Twenty trials make the denominator visible. They do not make the estimates precise enough to rank small differences. The paper acknowledges that individual contact-role ablations differ by only one or two successes in 20 trials, so their ordering is unresolved.

The contact bottleneck is visible at sub-goal level

The paper reports a diagnostic subset of 17 sub-goals from a 23-sub-goal suite. Every displayed free-space pick row reaches 100% for the full system. The weakest contact-critical sub-goal limits each final task chain:

• RAM alignment: 95%;
• CPU alignment: 85%;
• disk alignment: 95%;
• GPU alignment: 85%;
• first lever press: 50%.

This is stronger evidence than the 82% mean alone because it identifies where the sequence fails. It is still a selected 17-row diagnostic subset, not a complete publication of every sub-goal outcome and trial trace.

For more on stage reach and independent skill tests, see why one robot success rate can hide the failure.

Contact quality must remain separate from completion

A robot can complete an insertion by forcing the part through a bad alignment. A completion rate alone may reward that behavior.

Facet-0’s protocol defines several additional measures:

• peak off-axis force, the maximum force orthogonal to intended task-frame motion;
• force violation rate, trials with any step above a configured phase limit;
• intervention rate, trials in which a human supervisor takes over;
• recovery rate, detected failures resolved by the system;
• cycle time, wall-clock duration of a successful trial;
• inference latency, wall-clock time from observation to command.

One RAM ablation normalizes peak off-axis force to a force-free variant. Removing every predictive and evaluative wrench role cuts success from 95% to 45% and raises normalized peak off-axis force from 0.2 to 1.0. Individual component removals produce smaller changes that the 20-trial denominator cannot rank confidently.

The reported deployment-efficiency table gives 50 ms observation-to-command latency and 0.5 mm placement accuracy for the full system, compared with 150 ms and approximately 5 mm for π0.5. These are measurements in the authors’ system. They are not universal properties of the architecture.

A stronger publication would release the full distribution of force, torque, latency, cycle time, and violations, not only selected averages or normalized values. Tail events often decide whether contact is safe and whether parts survive repeated operation.

Recovery and human intervention are linked metrics

On a matched disk deployment corpus, the paper compares contact-selective post-training with advantage-weighted behavior cloning:

Metric	Matched AWR	Facet-0 post-training
Success	20%	65%
Human intervention	47%	24%
Recovery	44%	81%

The useful signal is the joint direction. Reported success and autonomous recovery rise while operator takeover falls. A success gain purchased through more human corrections would tell a different operational story.

The definitions still need careful reading. Intervention is reported per trial, while recovery is reported among detected failures. Neither denominator is interchangeable with interventions per hour, operator-control time, resets, or production uptime.

A complete recovery trace should identify:

1. what event triggered the failure detector;
2. whether the robot stopped before exceeding a declared limit;
3. whether it backed off, re-aligned, retried, replanned, or changed control mode;
4. how many retry cycles were allowed;
5. whether a human took control or only approved continuation;
6. whether the final part and fixture remained within quality limits;
7. whether the recovered attempt counts as one trial or a new trial.

Our robot intervention and handoff guide provides a denominator framework for those distinctions.

Adaptation is not free

Facet-0 uses a bounded local actor for part-specific adaptation while freezing the visual-language representation, reference policy, refinement expert, judge, compliant controller, and safety operator. Only 6.6% of policy parameters are updated in the unseen-module experiment.

The paper reports two adaptation views.

Few-shot transfer to an unseen module

For a memory module with different mass, stiffness, and latch response, every compared method receives ten demonstrations and three hours of training. Facet-0 reaches 45% success, while the strongest reported baseline reaches 5%.

That is evidence of a bounded transfer result, not immediate deployment readiness. A 45% final rate still represents more failures than successes under the stated test.

Adaptation dynamics on a known task

A separate RAM adaptation table uses a two-hour window and reports the 30-minute checkpoint. Facet-0 local adaptation reaches 95% success after 30 minutes, reaches 90% success in 10.5 minutes, and records a 5% violation rate across training episodes. The residual baseline reaches 80% at 30 minutes with a 20% violation rate.

These values make the training path visible, but publication should also retain:

• physical episode count and episode duration;
• supervisor-control time;
• number and type of interventions;
• failed and damaged parts;
• resets and calibration work;
• compute used during training;
• held-out final trials separated from training rollouts;
• regression tests on tasks the original policy already handled.

A method can update few parameters while consuming meaningful robot, operator, and fixture time. Our guide to the cost of rapid on-robot adaptation explains why demonstrations, wall-clock training, physical interaction, and retained capability need separate accounting.

What Peg-in-Bench adds

Facet-0 varies chassis pose and evaluates several electronics tasks, but its controlled suite remains one family of wrist-sensed, parallel-gripper assembly on a shared fixture. Peg-in-Bench addresses another evaluation need: repeatable changes in physical task configuration.

The proposed benchmark contains:

• five peg geometries: circular, rectangular, hexagonal, triangular, and L-shaped;
• three nominal clearance levels: 0.1 mm, 1 mm, and 3 mm;
• 15 shaped-hole pieces;
• eight orientations per shaped-hole piece;
• four connectable base variants;
• peg holders and mounting pieces;
• a seeded scenario generator that emits a visual layout and machine-readable JSON.

The authors recommend resin printing for pegs and holes at 0.05 mm printer resolution. They exclude a 0.02 mm clearance because common 3D printing could not reproduce it reliably. That exclusion is a useful benchmark-design decision: a nominal challenge is not reproducible if manufacturing variation is similar to or larger than the declared tolerance.

The framework targets three forms of generalization:

1. spatial generalization, new target positions and orientations;
2. geometric generalization, new shapes and clearances;
3. task-compositional generalization, multiple insertion operations joined into larger structures.

The first version intentionally holds illumination, camera placement, sensor noise, friction, material properties, and surface finish approximately constant. That isolation improves causal interpretation for task configuration. It also means the benchmark does not test robustness to those real-world changes.

A benchmark design is not a benchmark result

Peg-in-Bench does not publish a baseline policy table. Its future-work section calls for baseline results and standardized protocols. The paper demonstrates a physical setup with a UR7e and Robotiq gripper, but it does not establish a success rate, force profile, generalization gain, or winning policy.

The paper names a public repository for scenario generation and STL files. That URL returned HTTP 404 during our September 2 review. The artifact claim should therefore be treated as pending availability, not silently upgraded to a reproducible release.

A practical protocol for contact-rich evaluation

A lab, buyer, or reviewer can turn these lessons into a staged protocol.

1. Freeze the task contract

Declare:

• part and fixture versions;
• measured rather than nominal dimensions;
• clearance and angular tolerance;
• material, printer, post-processing, and wear state;
• initial pose distribution;
• success, quality, timeout, and abort conditions;
• configured force, torque, speed, and workspace limits.

Do this before trials. Do not tune the success rule after inspecting failures.

2. Separate familiar and unseen configurations

Create versioned training, validation, and test splits across:

• shape;
• clearance;
• position;
• orientation;
• material or stiffness;
• fixture;
• task composition.

A random seed is useful only if the scenario generator, component versions, and sampling distribution are also fixed.

3. Measure the complete control path

Report sensor exposure, preprocessing, policy inference, middleware, queueing, controller rate, and actuator command timing. Contact can change faster than a slow policy replans. A 50 ms model command does not replace the faster compliant-control loop beneath it.

The dynamic robot benchmark guide provides a fuller timing and process-quality protocol.

4. Publish contact outcomes beside task outcomes

For every trial, retain:

• final success and quality inspection;
• maximum axial and off-axis force;
• maximum torque;
• impulse or force-time integral where relevant;
• force-limit and safety-stop events;
• contact onset and duration;
• alignment and insertion depth;
• cycle time and stalls;
• part, tool, and fixture damage;
• intervention, reset, and recovery labels.

Do not use one composite score to hide a bad component. If a weighted decision score is needed, publish the underlying panels and weights.

5. Test controlled ablations

Useful ablations include:

• vision without force or tactile history;
• force input without force prediction;
• force prediction without a contact-aware value function;
• learned policy without local adaptation;
• local adaptation without the frozen reference;
• learned correction with and without compliant control;
• fixed geometry versus unseen geometry;
• familiar clearance versus tighter held-out clearance.

Ablations should preserve an executable system. Removing the low-level controller and causing immediate instability does not isolate the value of a learned representation.

6. Expose the adaptation ledger

Record all demonstrations, training episodes, interventions, violations, resets, exclusions, elapsed robot time, operator time, compute, and changed parameters. Evaluate on held-out trials after the stopping rule is reached, then rerun prior tasks to check retention.

7. Reproduce across laboratories

A meaningful outside reproduction should state:

• whether parts were downloaded or independently recreated;
• measured component dimensions and printer variation;
• robot, gripper, sensors, cameras, controller, and compute;
• calibration and task-frame procedure;
• software and model versions;
• scenario seeds;
• any deviations from the original protocol;
• trial-level results and confidence intervals.

Matching a nominal 0.1 mm clearance is not enough if two labs print materially different parts.

Evidence boundaries

Claim	Evidence class	Confidence	Boundary
Facet-0 reaches 82% mean success across five electronics assembly tasks	Author-run physical evaluation, 20 trials per method-task cell	High for the reported protocol	One shared fixture family, wrist sensing, parallel gripper, no independent replication
Contact failures are the main bottleneck in the evaluated suite	Reported task and selected sub-goal decomposition	Medium-high	Diagnostic subset contains 17 of 23 sub-goals
Joint action-wrench learning improves contact behavior	Matched recipe, baselines, and ablations	Medium-high	Several baselines are author reimplementations; small ablation differences are unresolved
Post-training improves recovery while reducing intervention	Matched disk deployment comparison	Medium	One task and author-defined detection, intervention, and recovery denominators
Ten demonstrations support transfer to an unseen module	Author-run few-shot adaptation study	Medium	Three hours of training, 45% final success, one unseen module
Peg-in-Bench can support reproducible task variation	Published benchmark design and machine-readable scenario specification	Medium	No baseline policy results; named repository unavailable during review
Either release proves production-ready contact manipulation	Not established	High confidence in the negative boundary	No customer uptime, throughput, quality, maintenance, safety certification, or independent field record

Questions to ask before trusting a precision result

1. What are the measured clearances, not only the nominal CAD values?
2. Which dimensions, materials, friction conditions, and fixture properties changed between training and test?
3. Does success require an acceptable force profile and quality inspection?
4. How many trials, seeds, parts, and physical setups support each percentage?
5. Are retries counted inside one attempt, and are reruns excluded?
6. Which force, torque, tactile, vision, and proprioceptive signals reach the policy?
7. Which layer commands motion, and which layer regulates force?
8. What are median and tail observation-to-command latencies?
9. How are jams, false seats, part damage, and protective stops classified?
10. What is the denominator for recovery and human intervention?
11. How much robot and operator time did local adaptation consume?
12. Did adaptation retain performance on prior tasks?
13. Are parts, task files, scenario seeds, code, checkpoints, and trial logs available?
14. Has another laboratory reproduced the result on measured physical artifacts?
15. What customer or production evidence exists beyond the lab protocol?

Verdict

Facet-0 offers unusually detailed physical evidence for a new contact-aware VLA recipe. It separates free-space and contact sub-goals, declares 20 trials per method-task cell, reports intervention and recovery, describes the controller beneath the learned policy, and publishes model, data, code, and project surfaces. Its 82% mean is meaningful inside that protocol.

The same evidence also defines the boundary. The evaluation remains author-run electronics assembly with one controlled fixture family. The hardest lever task succeeds in half of trials. The unseen-module result reaches 45% after ten demonstrations and three hours of training. No source establishes production uptime, certified safety, or outside replication.

Peg-in-Bench points toward the next step: make physical variation systematic and reproducible. Its modular geometries, clearances, orientations, structures, and seeded task descriptions are a useful design. Until baseline results and the named repository are available, it remains a benchmark proposal rather than an operating leaderboard.

The right conclusion is not that one model solved precision assembly. It is that contact-rich evaluation can be made inspectable. Publish the physical configuration, force quality, timing, learning cost, intervention, recovery, and artifacts beside every success rate.

Frequently asked questions

What is contact-rich robot manipulation?

Contact-rich manipulation is a task in which the forces created when a robot touches, slides against, inserts, presses, or fastens an object materially determine success. Vision may bring the tool near the target, but alignment, compliance, force, friction, and recovery decide whether the final interaction succeeds or jams.

Is insertion success enough to evaluate a contact-rich policy?

No. Report the number of trials, geometry, clearance, force limits, success rule, peak off-axis force, violations, intervention, recovery, cycle time, and excluded attempts. A robot can finish by applying damaging force, or fail safely after detecting a bad contact. Those outcomes should not be collapsed into one success rate.

What does Facet-0 actually prove?

In an author-run five-task electronics assembly suite with 20 trials per method-task cell, the fully adapted Facet-0 system averaged 82% task success, compared with 15% for the strongest reported matched baseline. The result supports the tested action-wrench representation, value-guided refinement, and local adaptation recipe on wrist-sensed parallel-gripper assembly. It does not establish independent replication, production uptime, certified safety, or general performance across every contact task.

Does Facet-0 command force directly?

No. The model predicts a future wrist-wrench profile as a consequence associated with a proposed Cartesian action. Only the Cartesian action is executable. A separate 200 Hz compliant controller closes its force loop on measured wrist force and torque.

How should adaptation cost be reported?

State demonstrations, physical episodes, wall-clock robot time, training time, parameter fraction updated, interventions, violations, resets, starting policy, stopping rule, final held-out trials, and retention on prior tasks. A few demonstrations can coexist with substantial supervised robot interaction.

What does Peg-in-Bench contribute?

Peg-in-Bench proposes reusable physical parts and a seeded scenario generator for varying peg shape, clearance, position, orientation, and assembly structure. It is a benchmark design rather than a policy leaderboard. Its paper does not publish baseline policy results, and the named repository was not available during our September 2 verification.